Detoxifying Large Language Models via Kahneman-Tversky Optimization

Currently, the application of Large Language Models (LLMs) faces significant security threats. Harmful questions and adversarial attack prompts can induce the LLMs to generate toxic responses. Therefore, detoxifying LLMs is a critical research topic to ensure their safe and widespread application. In this paper, we propose an alignment-based detoxification method for LLMs. We utilize Kahneman-Tversky Optimization (KTO) to align LLMs. During the construction of the training dataset, we take into account both the detoxification performance and the potential side effect on the LLMs. For detoxification, we make the LLM preferentially generate safe responses rather than toxic contents when asked with harmful questions and attack prompts. To mitigate the potential side effect on the conversational capabilities of LLMs, we incorporate normal questions into the training data, and ensure that the LLM generate normal answers, rather than safety refusals or unsafe responses. Experimental results show that our method showcase the best detoxification performance among all baseline methods while exerting little negative impact on the LLMs. Moreover, our method even enhance the LLMs' general abilities such as question answering and language understanding. Our proposed method achieve the first place in the NLPCC 2024 Share Task 10 Track 2 with an average score of 52.31.